Word Grammar. by Richard Hudson. Universität Tübingen, Word Grammar. Nika Strem, Iuliia Kocharina. Overview. The Cognitive Network

Transcription

1 by Richard Hudson Universität Tübingen, / 61

2 2 / 61

3 3 / 61

4 The Notion of Word grammar (WG) is a general theory of language structure WG is a branch of cognitive linguistics The main consideration is to be true to the facts of language structure WG assumption is that language can be analyzed and explained in the same way as other kinds of knowledge or behavior 4 / 61

5 Word as the Central Concept in WG The central unit of analysis is the word Grammar. Words are the only units of syntax, the largest and the smallest. Words provide the only point of contact between syntax and semantics Situation. Words are the basic units for contextual analysis Words are the nodes that hold the language part of the cognitive network together 5 / 61

6 Word as the Central Concept in WG 6 / 61

7 7 / 61

8 Language as Part of a General All knowledge is represented via a network of memory connections The sub-network of words is just a part of the total vast set of associations Thus, our language is a network of associations closely integrated with the rest of our knowledge 8 / 61

9 Language as Part of a General Traditionally, only the lexicon is viewed as a network in language (Aitchison 2012), storing only irregularities, while regularities are stored in a different way as rules (Pinker and Prince 1988). In WG view, exceptional and general patterns can both be accommodated in the same network due to default inheritance The only difference between these rules lies in two places: a verb versus come, and its stem followed by -ed versus came. Similarly, they can both be incorporated into the same network 9 / 61

10 Language as Part of a General 10 / 61

11 Language as Part of a General Vocabulary items are related in a network of phonological, syntactic and semantic links In language there is nothing but a network - no rules or principles or parameters or processes, except those that are expressed in terms of the network 11 / 61

12 Language as Part of a General The nodes of a WG network are atoms without any internal structure, so a language is not a network of complex information-packages such as lexical entries, constructions, schemas, or signs The information in each such package must be unpacked so that it can be integrated into the general network 12 / 61

13 Language as Part of a General 13 / 61

14 Labelled Links In cognitive psychology, there is a long tradition of associative network theories where all links have the same status In the WG view, links are classified and labelled stem, shape, sense, referent, subject, adjunct, etc. The classifying categories range from the most general (the isa link) to very specific categories 14 / 61

15 Labelled Links This approach allows named links to be used as functions, which yield a variable e. g. the referent of the subject of the verb defines one unique concept for each verb Every concept is uniquely defined by its links to other concepts, so labels are redundant. On the contrary, a network with unlabeled links is a mere associative network 15 / 61

16 Labelled Links Example: John saw Mary Labels serve to classify the links as same or different, so if we remove the label we lose information Each concept in a network is represented just once, with multiple links to other concepts rather than copies of the concept itself 16 / 61

17 Labelled Links 17 / 61

18 Modularity Is there a distinct module of the mind dedicated exclusively to language (or its layers)? In WG, the language network is a small part of a much larger network Language pathology: selective impairment of abilities Language network does not occupy a distinct part of the human mind or brain, but is intimately embedded in the general cognitive network (R. Hudson 2007) 18 / 61

19 19 / 61

20 Generalizations and exceptions Characteristics of a general category are inherited by instances of that category by default only when not overridden by a specific case (e. g. past tense) 20 / 61

21 is carried by the fundamental isa relation (e. g., because snores isa verb it automatically inherits all the known characteristics of a verb ) A fact is automatically blocked by any other fact which conflicts and is more specific 21 / 61

22 Multiple Multiple inheritance is an extension of default inheritance in which one concept inherits from two different concepts simultaneously: a cat inherits characteristics from mammal and from pet Ungrammaticality of *I aren t (You aren t vs. Aren t I?) due to conflicting inheritance from aren t (the negative present of BE) and am (the I-form of BE), without any way for either to override the other 22 / 61

23 Multiple Multiple default inheritance, defined by these facts about a concept A which isa B: : Normally A inherits all the characteristics of B and any other nodes on the isa chain leading up from B (i. e. any node which A transitive-isa) inheritance: But it does not inherit values for relations which already have a value Multiple inheritance: If A transitive-isa any other concept, it inherits from this in the same way as from B 23 / 61

24 24 / 61

25 Under WG, language is a network of concepts Language is part of the same general conceptual network that contains many concepts which are not part of language Language is a module in that the links among words are denser than those between words and other kinds of concepts, but it is not encapsulated or distinguished by special formal characteristics 25 / 61

26 A word may have a variety of links to other words and concepts: to its morphological structure via the stem and shape links to its semantics by the sense and referent links to its syntax by dependencies and word classes 26 / 61

27 27 / 61

28 Concepts are classified exclusively by means of inheritance hierarchies, which leaves no place for feature-descriptions: to classify a word as a verb in WG we give it an isa link to verb, not a feature-description which contains that of verb Multiple inheritance allows words to be classified on two different dimensions : as lexemes (DOG, LIKE, IF, etc.) and as inflections (plural, past, etc.) 28 / 61

29 29 / 61

30 The language network is a collection of words and word-parts (speech-sounds and morphemes) which are linked to each other and to the rest of cognition in a variety of ways, of which the most important is the isa relationship which classifies them and allows default inheritance 30 / 61

31 31 / 61

32 A WG analysis of an utterance is an extension of the permanent cognitive network, in which the relevant word tokens comprise a fringe of temporary concepts attached by isa links, so the utterance network has just the same formal characteristics as the permanent network A WG grammar can generate representations of actual utterances, in contrast with most other kinds of grammar which generate only idealized utterances or sentences 32 / 61

33 Word tokens must have different names from their types: identical labels imply identity of concept, whereas tokens and types are clearly distinct concepts. The WG convention is to reserve conventional names for types, with tokens labelled w1, w2, etc. Thus, I agree consists of w1 and w2, which isa I and AGREE:pres 33 / 61

34 Word types and tokens must have the same kinds of characteristics to allow inheritance. WG accommodates deviant input because the link between tokens and types is the Best Fit Principle (Miller 1992) 34 / 61

35 35 / 61

36 36 / 61

37 -free syntax Jo snores. NOT Jo tense snore. Exception: words that are both parts of other words and part of the syntactic structure Example: clitics Paul en mange beaucoup. Although the smallest units of syntax are always words, some words are not the smallest units of syntax but clusters of smaller words. Alternatively, we might distinguish phonological and syntactic words (Rosta 1997) 37 / 61

38 Inflectional morphology is responsible for mapping stems to shapes. Every lexeme has a stem, and by default a word s stem is also its shape Exceptions are due to inflectional morphology or to cliticization For example, BEHIND <behind>, DOG <dog> 38 / 61

39 39 / 61

40 40 / 61

41 In WG view, phrase structure is rejected In WG, syntactic structure consists of dependencies between pairs of single words A syntactic dependency is a relationship between two words that are connected by a syntactic rule The word-word dependencies form chains linking every word ultimately to the head of the phrase or sentence 41 / 61

42 42 / 61

43 A surface dependency analysis can always be translated into a phrase structure Dependency analysis is much more restrictive than phrase-structure analysis The extra richness of dependency analysis lies in the labelled dependency links, as well as in the possibility of multiple dependencies 43 / 61

44 Structure-sharing Structure-sharing is found when one word depends on more than one other word - i.e. when it is shared as a dependent It is the main device in WG which allows phrases to be discontinuous 44 / 61

45 45 / 61

46 In a surface structure, each dependency is licensed by the grammar network In case of structure-sharing, just one of the dependencies is drawn above the words Any of the competing dependencies could be chosen, but only one choice satisfies a well-formed surface structure (free of tangling and dangling ) The totality of dependencies constitutes the surface structure 46 / 61

47 Coordination Conjuncts must share their external dependencies The structure of the coordination itself ( conjuncts and coordinators ) is analyzed using word-strings simple strings of words built up by ordinary dependencies 47 / 61

48 48 / 61

49 49 / 61

50 WG has a compositional semantics in which each word in a sentence contributes some structure stored as its meaning These meanings are concepts defined by a network of links to other concepts: APPLE and PEAR EAT 50 / 61

51 Words of all word classes contribute the same kind of semantic structure, which in WG is divided into sense (general categories) and referent (the most specific individual or category referred to) similar to the contrast in morphology between stem and shape 51 / 61

52 52 / 61

53 The way in which the meanings of the words in a sentence are combined is guided by the syntax, but the semantic links are provided by the senses themselves 53 / 61

54 54 / 61

55 A word s basic sense (the one inherited from its lexeme) is modified by the word s dependents, which produces a second sense, more specific than the basic sense but more general than the referent This intermediate sense contains the meaning of the head word plus its dependent, so in effect it is the meaning of that phrase 55 / 61

56 In contrast with the syntax, the semantic structure contains a node for each phrase, as well as nodes for the individual words i. e., a phrase structure The step-wise composition of word meanings is called semantic phrasing 56 / 61

57 57 / 61

58 The advantage of WG for a parser is the lack of invisible words, but the dependency basis also helps by allowing each incoming word to be integrated with the words already processed without the need to build (or rebuild) higher syntactic nodes A simple algorithm guides the search for dependencies in a way that guarantees a well-formed surface structure 58 / 61

59 Short sentences make good examples. 1. w1 = short. No progress:- w1. 2. w2 = sentences. Capture:- w1 w2, 3. w3 = make. Capture:- W1 6 w2 w3, 4. w4 = good. No progress:- w1 w2 w3, w4. 5. w5 = examples, Capture:- w4 e w5, 6. Submit:- w1 w2 w3 (w4 ) w5 59 / 61

60 Hudson s theory of syntactic complexity builds on this incremental parsing model The aim of the parser is to link each word as a dependent to some other word This link can most easily be established while both words are still active in working memory Once a word has become inactive it can be reconstructed (but this is costly) 60 / 61

61 Thus, short links are always preferred to long ones, which enables easy calculation of the processing load for a sentence: the mean dependency distance (the number of other words between a word and the word on which it depends) In English texts, dependency links tend to be very short (70% of words are adjacent to the word on which they depend, with 10% variation in either direction) 61 / 61

62 Aitchison, Jean (2012). Words in the mind: An introduction to the mental lexicon. John Wiley & Sons. Brown, Dunstan et al. (1996). Russian noun stress and network morphology. In: Linguistics 34.1, pp Covington, Michael A (2001). A fundamental algorithm for dependency parsing. In: Proceedings of the 39th annual ACM southeast conference. Citeseer, pp Fraser, Norman M (1989). Parsing and dependency grammar. In: Carston, R obyn (ed.) UCL W or k-in g P a pers in L in g uistics 1, p. 29. Fraser, Norman M and Richard A Hudson (1992). in word grammar. In: Computational Linguistics 18.2, pp Hudson, Richard (2010). An introduction to word grammar. Cambridge University Press. (2007). Language networks: the new. Oxford University Press. 61 / 61

63 Hudson, Richard (1998). Word grammar. Miller, JE (1992). Richard Hudson, English. Oxford: Basil Blackwell, Pp In: Journal of Linguistics 28.02, pp Pinker, Steven and Alan Prince (1988). On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. In: Cognition 28.1, pp Rosta, Andrew (1997). English syntax and theory. PhD thesis. University College London (University of London). 61 / 61